Q1. What is fundamentally different about deploying agentic AI versus traditional AI or copilots within financial services? Where do you see the biggest opportunities for agents to transform front-office, middle-office and back-office workflows?

The industry has moved through three distinct stages: single-turn chat assistants, then retrieval-augmented (RAG) document Q&A, and now agentic systems. The difference is not incremental. A copilot compresses a task a human is already performing. Agentic ecosystems own workflows and entire business processes, with humans in the loop gating the decisions that matter: they reason through ambiguity, construct multi-step plans, invoke tools and enterprise APIs, spawn sub-agents, and self-correct from runtime feedback.

That autonomy changes the risk surface entirely. A copilot's error is caught by the human in the seat. An agent's error propagates — into the next tool call, the next agent, the next decision. Some early adopters of AI are calling it Asset Management 2.0 — defined not by faster information retrieval but by AI integrated directly into the core investment workflow. That depth of integration is exactly why the discipline has to change.

On opportunity, the aggregate number is striking — industry expectations of 30–50% efficiency gains from an average asset manager's cost base are not far fetched. But the distribution matters more than the headline:

Front office. Investment research review is the strongest early candidate: agents that ingest internal and sell-side research, extract theses, weight evidence by source track record, surface contradictions, and route everything through a human gate before anything touches a recommended list. High analyst leverage, comparatively low regulatory exposure, fast signal on whether the architecture works. The critical constraint, which production deployments must internalize: agents must attach source references to every insight they surface — provenance is not optional with AI.

Middle office. Client mandate application is underrated. Parsing investment policy statements, classifying hard versus soft constraints, translating them into portfolio rules, and logging every interpretation choice is exactly the kind of expert reasoning workflow agents handle well. The same shape repeats across the middle office: investment guideline monitoring, where an agent detects a potential pre- or post-trade breach, assembles the position history and the relevant mandate language, and escalates with the evidence pack ready for the compliance officer; performance attribution and client reporting, where agents draft commentary grounded in the attribution data and flag anomalies against prior periods; and corporate actions, where routine elections process straight through while complex or economically material events route to a human with the analysis already attached.

Back office. Exception handling, reconciliation breaks, settlement fails — event classification and routing with full context. Routine in-policy flows automate; exceptions escalate to humans with the evidence already assembled.

There are places where agents don't belong: anything inside a hard latency budget — the execution hot path — is not an agent problem. Agents belong in the deliberative loop around trading (pre-trade strategy selection, post-trade TCA), not in the microsecond loop.

Q2. As financial institutions move from AI experimentation towards deploying autonomous agents, how should firms think about data readiness for AI agents? What needs to change in terms of data quality, accessibility, lineage and context for agents to make reliable decisions?

The uncomfortable truth is that most firms are trying to deploy agents on data estates that humans already struggle to navigate. An agent doesn't fix that problem — it amplifies it, at machine speed and machine volume.

Start with the scale of the existing waste. Several industry studies have reported that data users spend in excess of a third of their time in searching for data where no clear data inventory or registry is available, and then more time evaluating and/or cleansing the data to ensure it is in a form and context for their use case. That works when the consumers are patient humans who can ask a colleague what a field means. It is fatal when the consumer is an autonomous agent that will happily execute a multi-step plan on top of a misinterpreted column.

So data readiness for agents is less about pristine quality scores and more about machine-consumable meaning. Four things have to change:

First, a semantic layer becomes non-negotiable. Every business term — "client," "position," "exposure," "fee" — needs a single authoritative definition bound to the physical data that implements it. In my own framework for valuing this investment, I've argued that each semantic assertion should be bound to a specific physical column with provenance and freshness metadata attached — otherwise the ontology rots and agents inherit stale meaning.

Second, lineage becomes a runtime dependency, not a compliance artifact. When an agent produces an output that informs an allocation or a client report, "where did this number come from?" must be answerable programmatically, in the moment — not reconstructed weeks later for an examiner.

Third, context must be captured from the people who hold it. In most firms, the real meaning of data lives in the heads of senior operators — tribal knowledge. The practical path I've seen work is incremental, crowdsourced capture: business users annotating fields in small increments inside the applications they already use, with those annotations accumulating into an enterprise semantic layer. This makes data AI-ready without a multi-year migration.

Fourth, entitlements need to be rethought for non-human consumers. Buy-side data contracts were written for named human users. An agent fleet querying licensed market data raises genuine questions about scope of use that firms should resolve with their vendors before, not after, production deployment.

Q3. Many financial institutions have successfully experimented with AI but struggle to move into production at scale. What are the biggest technology, data or organisational barriers preventing agentic AI from moving beyond the proof-of-concept stage?

The dominant failure pattern is horizontal: broad pilots — generic chatbots, summarization tools, document Q&A — that generate activity and demos but never compound into operating-model change. Every competitor runs the same pilots on the same commodity models and gets the same temporary lift. Productivity gains from generic deployments reset the floor of the industry; they don't move anyone's ceiling.

Beneath that pattern sit four specific barriers:

The data-readiness gap. As covered above — agents deployed on data without semantic context, lineage, or provenance produce outputs no fiduciary can stand behind. Firms discover this at the compliance review, which is the most expensive possible place to discover it.

The observability and governance void. Pilots run without production-grade telemetry, and each team's framework keeps its logs in a silo. When risk and compliance ask the reasonable questions — what did it do, why, at what cost, with whose approval — the pilot team has no answer, and the initiative stalls at exactly the moment it needed institutional trust. This barrier is large enough, and solvable enough, that it deserves its own discussion.

Rebuilding the plumbing per use case. When every use case re-implements validation, human-gating, audit, and memory from scratch, use case number five costs as much as use case number one and the economics never turn. The fix is architectural: build the domain-agnostic layer — validation, human-in-the-loop framework, institutional memory, structured reasoning phases — once, and let every workflow inherit it. Then each new workflow is a variable-cost addition against a fixed-cost backbone, and AI becomes operating leverage rather than an expense line. This is not a theoretical pattern — this is the Operating Model of Asset Management 2.0.

The organizational barrier — and it's the biggest. Compliance and risk engaged as end-of-pipeline approvers instead of co-designers. Every production blocker they raise late was knowable early. Inverting that relationship — validation acceptance criteria and audit standards agreed at design time — is the single highest-leverage organizational change available.

There's also a scope discipline the industry keeps relearning: narrow the MVP until the timeline is defensible. One workflow, end-to-end, on real production data, under real fiduciary and audit conditions, in a quarter — proven, then extended. Multi-workflow moonshot programs sound strategic and deliver very little against expectations. Reusable accelerators — pre-built integration patterns, agent scaffolding, evaluation harnesses carried from engagement to engagement — are what compress the time after the first MVP.